Chen Zeng

𝗧𝗵𝗲 𝗯𝗲𝘀𝘁 𝗔𝗜 𝗰𝗼𝗱𝗶𝗻𝗴 𝗹𝗲𝘀𝘀𝗼𝗻 𝗰𝗼𝘀𝘁 𝗼𝘂𝗿 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿 $𝟲𝟬𝟬 𝗮𝗻𝗱 𝗮 𝗺𝗮𝗿𝗿𝗶𝗮𝗴𝗲 𝗮𝗿𝗴𝘂𝗺𝗲𝗻𝘁. Our VP of Engineering, Xiaofan(James) Luan, was supposed to buy his wife a Dior bag for their anniversary.  Instead, he bought three Claude Code subscriptions and spent the holiday trying to cross-compile 2 million lines of C++. Every fix on one platform broke two others.  $600 later, the only output was "git reset --hard" — and a very cold dinner table.😂 "Make it compile on Windows" is a trap. The real goal was "compile everywhere without hacks" — no AI is going to figure that out for you at 2 am. What worked: constraints before code, review tests not code, bottom-up, one layer at a time. Same task, two days. Then he ran six parallel Claude sessions across three machines with git worktree. The bottleneck stopped being intelligence and started being how fast one person can alt-tab. AI solves exactly the problem you give it. Engineering is in knowing which one to give. His wife is still waiting for that bag. Full story: https://lnkd.in/gtsW_Wvk ——— Follow Milvus, created by Zilliz, for everything related to unstructured data

9

Google Research Director Denny Zhou, who founded the LLM Reasoning Team at DeepMind, recently gave a great talk at Stanford’s CS25 class: Large Language Model Reasoning. This was one of the most intuitive talks about reasoning model training and application. He outlined new directions for training models to handle questions where answers aren’t easily verifiable. My team is currently working on an enterprise LLM project focused on using reasoning to extract relevant context and generate responses for unverifiable cases. It’s a tough but promising area, and it’s an interesting time to be building. [Link to Denny Zhou's talk: https://lnkd.in/gdJ6Mzi8]

1,026

8 Comments

We (w/ Shansan Gong, Ruixiang ZHANG, Huangjie Zheng, Jiatao Gu, Navdeep Jaitly, Lingpeng Kong) released a family of 7B diffusion language models, DiffuCoder, that specializes on code generation, with a focus on understanding and improving masked diffusion models. A core analysis of DiffuCoder is the autoregressiveness (AR-ness) score, a novel metric that quantifies the causal patterns in decoding, revealing how diffusion models break from strict left-to-right generation for more flexible, non-linear code planning. Recent advances in autoregressive (AR) models dominate code generation, but diffusion-based LLMs (dLLMs) like DiffuCoder offer a promising alternative, especially for complex programming tasks. DiffuCoder explores how these models decode differently—showing less global AR-ness in code tasks compared to math—and how temperature affects both token selection and generation order, unlike traditional AR models. We also introduce coupled-GRPO, a post-training RL method with a coupled-sampling scheme, to reduce performance drops during accelerated decoding, boosting parallelism and efficiency. We use a self-improvement pipeline that leverages AR-ness analysis, coupled-GRPO optimization, and evaluation on benchmarks like AceCode-89k to refine decoding strategies. This approach enables DiffuCoder to navigate diverse code generation pathways and enhance performance with modest computational overhead. Looking ahead, we aim to further leverage Reinforcement Learning to steer code generation through these decoding patterns, with the discrete nature of AR-ness scores providing a foundation for search-based strategies—ideal for the sparse rewards of optimizing complex code structures. Check out our full paper and code for a deeper dive! Paper: https://lnkd.in/gVWU3BDJ Code: https://lnkd.in/gmXTZ_6n Models: https://lnkd.in/gTcKCDr9 #MachineLearning #AI #CodeGeneration #DiffusionModels #NLP

220

5 Comments

In 6 hours, I helped my friend Yasmin Primrose build what an external firm quoted as ~20 FTE days of Azure Batch infrastructure work — and we spent at least half that time waiting (permissions, quotas, RBAC). She built a complete end-to-end Azure Batch pipeline: SimPy containers built and deployed, batch pools provisioned, blob storage for input and output, and a way to create and monitor jobs. Lesson: the bottleneck wasn't writing code. It was getting permissions, quotas, resources, and meeting governance standards. The final boss? Azure role-based access controls. A sampling of the challenges. - UK South: out of vCPUs, no batch pools - Sweden Central: vCPUs available, but batch pool quota set to zero - Lack of subscription owner permissions - Have permissions, but elevated roles had already expired Yaz was helping a major retailer run supply-chain simulations that were taking hours or days on a laptop. She wanted to fire dozens of jobs into the cloud so the team could iterate on "what if" questions fast enough to actually matter. Some sessions were amazing, building out infrastructure at lightning speed. Yaz said that it felt like "breaking through an invisible barrier." In contrast, our third session was nothing but barriers (ha!), mostly beyond our control. Thankfully, the person in charge of security was on-call and joined our session, and could see the problems we were having. He was opening support tickets to Microsoft and responding to their questions in real-time. The idea of having security on-call during pairing sessions came from Ben Grinnell FBCS (Board Member, Capabilities and AI) at Newton. What a fantastic idea—if you know you'll get stuck on RBAC/quotas, make sure you unblock it in real time. A group CTO from a private equity firm said problems like these will pose the greatest leadership challenge of the last 100 years. That's a bold claim. But in 15 minutes, he convinced me. His argument: every two-day delay used to be tolerable because it was just one of a hundred things that each delayed us two days. Now that coding is instantaneous, those delays are all that's left. It's literally all wait time. I'm so delighted that Ben will be co-presenting with Tom Kilcommons (AI Innovation Lead) from Newton will present on their novel AI practices to accelerate creating client value at ETLS San Jose, April 9–10. If you're wrestling with problems like this, you'll see concrete patterns from leaders at OpenAI, Amazon, Disney, and Netflix. Full program below — see you there! https://lnkd.in/gxZukzA6

34

1 Comment

After the release of iOS 26, I feel confident that AI is not replacing senior engineers any time soon. The progress of LLMs has been remarkable. Scaling compute and data has unlocked surprising capability. There are many well understood tasks that models can now handle, and that will continue to reshape the jobs market. But senior engineers spend most of their time on work that is not well understood. They operate in ambiguity. They make sense of new platforms, new APIs, new constraints. They combine broad experience with intuition to navigate the unknown. Most importantly, their context window grows over years when focused. Over the past few months I used ChatGPT 5 and Claude Sonnet 4.5 inside Xcode while updating Land of Livia to iOS 26. Both often sent me down dead ends with the confidence of a sociopath. Edge cases, platform inconsistencies, silent failures. The models could not reason about a new system based on limited training data. It’s not a matter of scaling data centers. That kind of reasoning requires approaches beyond LLMs. Until then, the frontier still belongs to humans.

42

7 Comments

just talked to a hardware startup in sf as a team of 2, that just raised $M for their seed round. they've hit: hundreds or pre-orders, millions of impressions on social, and 2nd iteration of product in less than 10 months. ⚡ speed is the most important thing for building. and we're excited to help anyone who wants to test their speed this summer in ai + hardware with the 30 worldwide hack we're hosting with Boost VC, @Flux, Studio 45, and more. we treat building like a sport at worldwide studios, and think there needs to be more open access sprints like this. more info below.

47

1 Comment